Every AI adoption story seems to start the same way: excitement, a demo, a signed contract, and then, months later, a nagging sense that nothing much has changed. In a conversation on Isolio's podcast, Tamas Feher, founder and CEO of Isolio, sat down with IT consultant Akos Voros, whose two decades of experience stretch from IBM to ELTE's Applied AI Center, to talk about why so many small and mid-sized businesses end up disappointed by AI, and what separates the companies that get real value from the ones that just get a bigger software bill.
Voros's answer starts before any tool is chosen.
Start at "point zero"
He calls it "point zero": before a business talks about transformation at all, it needs to know what business benefit it is actually expecting from AI, expressed in concrete, measurable terms. Not enthusiasm, not fear of missing out, but a number. When will this investment turn around and start generating money for the company? If the plan is to free up staff time, does the business already know what that freed-up time will be used for?
He's blunt about why this matters. Every technology wave produces the same pattern: a period of hype, a lot of enthusiasm, and then a wave of disappointed communities forming online to compare notes on what didn't work. He's watched it happen with prior tech cycles, and he sees the early signs of it happening again with AI, from groups full of frustrated chatbot builders to the broader graveyard of over-promised automation projects.
His advice is to skip that cycle entirely by refusing to sign a contract, or start a project, without first agreeing what "success" actually looks like in currency terms, and over what time horizon.
What actually belongs to a machine
Once a business has that measurable target, the next question is which tasks to hand over. Voros's rule of thumb is refreshingly unglamorous: give AI the work that people don't enjoy doing anyway, the tasks that are repetitive and tiring but can be described clearly with rules.
That combination, boring plus rule-describable, is what keeps risk low. If a task can be specified precisely, there's less room for an AI system to go off the rails, hallucinate, or make a judgment call it isn't equipped to make.
This is also where he introduces a framing that a lot of the current AI discourse skips over: think of AI systems as digital colleagues who need a job description, the same way a new human hire does. That job description should spell out exactly what the role covers and, just as importantly, what it doesn't.
The value isn't the hours saved
The conversation in most companies stops at the headline number, something like "we saved so many hours by having an agent handle invoicing." What's missing, Voros argues, is the next question.
An agent might realistically handle seventy or eighty percent of invoicing, because there will always be edge cases it can't resolve. If that saves a member of the finance team two or three hours a day, the real value isn't the hours saved by the software. It's what that person does with the two or three hours they get back.
Often, that reclaimed time turns out to be worth more than the automation itself, because it lets someone actually sit down and evaluate whether a given vendor contract is still the best deal available, or whether it's worth renegotiating. That kind of judgment work is where the value gets created, not in the automation step itself.
Chasing the last twenty percent will bankrupt the project
Both participants push back on a specific kind of ambition: the urge to automate a process completely. Voros's advice is almost the opposite of what a lot of AI vendors are selling. Don't aim to eliminate a hundred percent of a workflow's manual steps. Aim for something like sixty to eighty percent.
Those first steps are the ones that can genuinely be described with clear rules, and the system built around them will be dramatically simpler and cheaper than one that tries to swallow the entire process, including every unpredictable edge case. He compares over-ambitious full automation to Darth Vader knocking on the door with the Death Star's business plan: technically complete, but nobody actually knows how to build the whole thing yet, and trying will eat a month of effort chasing a result nobody can fully specify.
Start with the achievable, healthy majority of a process. Once that's running, the patterns that emerge make it much easier to figure out how to extend automation further, one rational step at a time. The remaining fraction tends to obey something close to a Pareto pattern: a shrinking share of the work absorbs a disproportionate share of the effort to automate, and often it's smarter to just leave it with a person.
AI didn't take anyone's job. People did.
Perhaps the sharpest point in the conversation is about layoffs. Both describe watching companies lay off employees in the name of AI transformation, and then quietly rehire once it became clear the technology couldn't fully cover the role.
Tamas makes the case plainly: the popular narrative that "AI took the jobs" gets the causality backwards. What actually happened is that people made staffing decisions based on hype rather than evidence, on the assumption that AI could fully replace a role when the technology, at least in its current form, generally can't. It wasn't AI that laid off those employees. It was managers acting on incomplete information.
Voros brings in a concrete example from outside the software world. In 2016, prominent predictions suggested that AI would make the job of reading medical scans largely obsolete within five or six years. What actually happened is that more people work in radiology today than before, and they're paid more, not less.
The explanation is the one that runs through the whole conversation: even when AI can competently handle ninety percent of a task, the remaining ten percent doesn't disappear. It often becomes more valuable, because the judgment, oversight and accountability that remain squarely human get concentrated into a smaller, higher-stakes slice of the job.
There's also a practical wrinkle worth remembering: it might be technically possible to automate ninety percent of a workflow, but if that ninety percent is expensive to run at scale, a business may be better off automating a more modest thirty or forty percent, specifically the repetitive parts that staff turnover shows people don't want to do anyway. That cost question is the subject of the second article in this series.
The upgrade, not the replacement
That reframing leads to the most optimistic thread in the discussion. If AI absorbs the tedious, rule-bound parts of a role, the humans in that role don't necessarily do less work; they get pushed toward higher-value activity.
Tamas points to a pattern he hears about constantly: employees currently spend real chunks of their day on manual admin, copying data from one system into another, when what the business actually wants is for them to spend that time with customers, building relationships and the brand. Automating the copy-paste work doesn't eliminate the human role. It upgrades it into something more interesting and, both agree, probably more satisfying too, while the harder judgment calls the machine still can't reliably make stay with people for the foreseeable future.
The conversation closes on a line that frames the whole episode: the future isn't automatically good or automatically bad. It depends entirely on how individual companies, and society more broadly, choose to prepare for it. Get the measurable "point zero" right, choose the right slice of work to hand over, resist the urge to automate everything at once, and the odds of landing on the good side of that outcome go up considerably.
The next two articles in this series take that further: why today's AI pricing won't hold, and how to cap the cost and scope of an agent before it surprises you.
This article is drawn from a conversation on Isolio's podcast between Tamas Feher (Isolio) and Akos Voros, an IT consultant whose career spans IBM and ELTE's Applied AI Center.